Refugees in Kenya are powering tech for dwindling pay

In the Kakuma refugee camp in northern Kenya, near the border with South Sudan, refugees do remote gig work such as content moderation and data annotation for tech companies. This kind of work was once promoted by UN agencies and global tech companies as a win-win: refugees get income opportunities, and AI and social media companies get cheap labor.

But The Guardian interviewed more than 20 refugees in Kakuma and concluded that only one side clearly benefits from this arrangement. Tech companies used datasets processed by precarious workers, including refugees, to build their most powerful tools; and the refugees' pay, already low, now seems to be evaporating as entry-level tech work disappears.

A refugee using the pseudonym Grace is one example. She uses ChatGPT to research Christian hymn translations in a language she does not speak, translating English text into an assigned East Asian language, then using the translation results to search for existing versions and look up information about translators, authors and denominations, and finally submitting these details to AOP Connect. She does not know who the client is, what the purpose of the work is, or whether she will be paid.

AOP Connect is a crowdsourced research platform under RWS, a British AI enterprise software company. According to RWS researcher documents and manuals reviewed by The Guardian, if RWS judges the work to be high quality under six criteria, it can grant a "discretionary reward" of up to $500.

The problem is that refugees are not guaranteed wages. Asha Luka Abdallah, a mentor at SIR (the Solidarity Initiative for Refugees), said most people receive far less than the maximum, and no one tells them why. RWS says higher-quality work brings higher pay, but according to project participants and payment notices reviewed by The Guardian, there is almost no transparency about how evaluations are conducted.

On AOP Connect, Grace saw that the hymn research project was listed with a $9,000 reward pool. But that amount is divided among an unknown number of contributors, and it is a discretionary "reward" rather than a guaranteed wage — a point already stated in the project listing's description.

An AOP Connect spokesperson responded that the platform "is not AI-driven, nor is it used for AI model training work," that the reward structure is communicated "transparently in advance," and that pay is based on the quality, relevance and timeliness of submitted research. The spokesperson also said confidentiality is crucial to AOP's clients. But The Guardian's report provides no independent verification, and whether the platform is actually used for AI model training remains in doubt.

Kakuma was not always like this. It was originally established as a temporary home for people fleeing the Second Sudanese Civil War, and later took in people from Somalia, Ethiopia and the Democratic Republic of the Congo. More than 30 years later, it has become one of the world's largest refugee camps, with rows of tin-roofed buildings that make it look more like a permanent town.

The area around the camp is arid and severely short of water, agriculture is not viable, and more than 300,000 residents depend almost entirely on humanitarian aid. In a place like this, any income has value: fewer than half of households have income, and those that do earn less than $50 a month.

A few years ago, SIR staff said they could easily find data annotation, transcription and translation work for students on platforms such as Remotasks. SIR has trained more than 2,000 refugees and connected more than half of them to remote tech work, ranging from long-term employment to short-term gigs.

But Scale AI subsidiary Remotasks exited Kenya in 2024. SIR's Hydrogene said that before Remotasks shut down, workers typically took home 500 to 1,000 Kenyan shillings a day, about $3 to $6. The income was low, but in Kakuma it was hard to come by. After the channels narrowed, some people had to accept work that substitutes potential "rewards" for guaranteed wages.

For tech workers in Kakuma, the bad news does not end with the platform's exit. A South Sudanese software engineer said that because of AI, entry-level work is disappearing, and senior engineers no longer hire interns or junior positions. He himself obtained well-paid work through SIR and the French organization Konexio, but he worries about future workers.

This observation fits a broader picture: tech companies build tools using datasets processed by precarious workers, and those tools are in turn replacing entry-level tech work. But the report provides no quantitative evidence that AI is replacing entry-level work; the interviewee's personal observation can only serve as a signal of a trend, not a definitive conclusion.

The Nairobi-based Data Labelers Association calls these workers the tech industry's "invisible architects," arguing that they sit within an opaque supply chain. The association's founding chair, Joan Kinyua, has done much of this work herself, and she points out that the micro-work economy is often exploitative.

There is a key constraint in Grace's situation: as a refugee who cannot legally work in Kenya, she needs any opportunity she can get. This explains why she still took on the hymn research project without knowing the client, the purpose or the pay.

Work on AOP Connect is bound by confidentiality agreements, and workers may not disclose clients or work content. Grace requested anonymity because of the confidentiality agreement. This confidentiality arrangement, layered on top of the opaque reward mechanism, makes it hard for workers to judge what their labor is worth and hard for them to organize for better conditions.

The report does not provide the full extent or time trend of the "continuous shrinking" of refugee gig pay, nor systematic data on the actual income and job stability of refugees whom SIR connected to remote work. These gaps are themselves a limitation of the reporting and a reflection of the state of this issue: changes in the pay of the workers at the very bottom often lack trackable data.